Abstract: Environmental pollution caused by garbage is a significant problem in most developing countries. Proper garbage waste processing, management, and recycling are crucial for both ecological and economic reasons. Computer vision techniques have shown advanced capabilities in various applications, including object detection and classification. In this study, we conducted an extensive review of the use of artificial intelligence for garbage processing and management. However, a major limitation in this field is the lack of datasets containing top-view images of garbage. We introduce a new dataset named “KACHARA,” containing 4727 images categorized into seven classes: clothes, decomposable (organic waste), glass, metal, paper, plastic, and wood. Importantly, the dataset exhibits a moderate imbalance, mirroring the distribution of real-world garbage, which is crucial for training accurate classification models. For classification, we utilize transfer learning with the well-known deep learning model MobileNetV3-Large, where the top layers are fine-tuned to enhance performance. We achieved a classification accuracy of 94.37% and also evaluated performance using precision, recall, F1-score, and confusion matrix. These results demonstrate the model’s strong generalization in aerial/top-view garbage classification. PubDate: Wed, 02 Jul 2025 03:19:16 +000
Abstract: A traffic network exhibits inherent characteristics of networks while also possessing unique features that hold significant research value. In this study, the limitations of static graph structures and the challenges of accurately modeling spatiotemporal dependencies in traffic flow have been addressed through a hybrid GCN-gated recurrent unit (GRU)-transformer model. The proposed model integrates a dynamic topology module (DTM) with graph attention networks (GATs), GRUs, and transformer-based temporal modules to adaptively capture the evolving dynamics of traffic networks. The DTM dynamically updates graph structures based on real-time traffic conditions, while GATs focus on identifying critical spatial relationships. GRUs efficiently capture temporal dependencies, and the transformer model captures long-term sequential patterns, providing a comprehensive framework for real-time traffic forecasting. The proposed model was trained and evaluated using the METR-LA dataset, which comprises traffic data from 207 sensors at 5-minute intervals. The model demonstrated superior performance across various metrics, achieving RMSE, MAE, and MAPE values of 4.125%, 2.985%, and 5.432%, respectively, for 15-minute predictions, with an value of 0.928. For longer prediction horizons (30, 45, and 60 min), the model consistently outperformed baseline methods, maintaining competitive RMSE and MAPE values. The experimental setup included normalization, graph construction using adjacency matrices, and preprocessing steps to ensure data quality and robustness. The integration of spatial and temporal features through the GCN-GRU-transformer framework enhanced the model’s ability to generalize across varying traffic scenarios, including peak hours and disruptions. Compared to traditional methods, which often rely on static graphs and fail to adapt to real-time changes, the hybrid model effectively addresses both spatial heterogeneity and temporal dependencies. The results indicate its robustness in handling complex traffic dynamics, adaptability to real-world variations, and potential applications in intelligent transportation systems. Future work will focus on incorporating multimodal data sources and enhancing computational efficiency to achieve broader scalability and deployment in smart city infrastructures. PubDate: Thu, 26 Jun 2025 04:04:59 +000
Abstract: Medical imaging plays a crucial role in modern healthcare, facilitating the diagnosis and treatment of various diseases. The advent of deep learning has revolutionized the processing and analysis of medical images. This paper reviews recent literature on deep learning applications in medical imaging, focusing specifically on segmentation and classification for disease diagnosis and treatment. We discuss recent advancements in deep learning architectures tailored for these tasks, highlighting their relevance and effectiveness. The studies reviewed span the period from January 2023 to April 2024, concentrating on the latest deep learning methods proposed for breast cancer segmentation. Additionally, we explore the availability and characteristics of publicly available medical image datasets for breast cancer, emphasizing their importance in training and evaluating deep learning models. An overview of commonly used metrics for assessing model efficacy is provided, underscoring their role in quantifying performance. Furthermore, we address the challenges and limitations faced by deep learning methods in medical imaging. Through analysis and discussion, we propose innovative directions to address these challenges, paving the way for promising future applications in early disease detection and personalized treatment planning. PubDate: Mon, 23 Jun 2025 20:14:00 +000
Abstract: The prevailing method for grading oil palm fruit bunches in mills relies on human graders conducting visual inspections, resulting in frequent errors and inconsistent assessments. This is a significant open problem when developing a detector for oil palm fruit ripeness and oil content, which are related factors. Current trends focus on computer vision techniques based on image processing and machine learning to improve grading of oil palm fruit bunches at an individual factory, resulting in limited diversity of the data used for evaluation. Collecting data from all factories offers informational advantages but raises privacy concerns. Addressing these challenges, this study proposes federated learning (FL) to develop local and global prediction models for grading oil palm fruit ripeness while preserving data privacy. FL facilitates collaborative model training across factories, mitigating privacy risks and enhancing model development efficiency. The proposed model uses the color of palm husks to determine the ripeness stage, which is used as a factor in predicting the yield of oil from the crop. A predictive model was created using FL principles with a training dataset of 5209 images, which was divided into two subsets: single-palm (2571 images) and multipalm (2638 images). The classification accuracy of a global model was 90.0%, while the local models were expanded to include private data for each of 4 testing clients. The predictive global and local models from FL were used to implement the system in a web application form and to validate its performance in calling the oil palm ripeness stage. PubDate: Sat, 21 Jun 2025 00:05:39 +000
Abstract: Meningioma, a common type of brain tumor, highlights the critical need for early detection and treatment. The proposed research work addresses the severity level of meningial brain tumors by classifying the meningioma grades. The proposed work makes use of a dataset containing 35 Grade II and 49 Grade I images. To achieve accurate binary classification of meningioma grades, two deep learning models SimpleCNN and VGG16 + XGBoost were proposed. The proposed models use MRI image dataset to train the SimpleCNN model, allowing it to directly extract pertinent characteristics from the images. VGG16 model is a pretrained model used for extracting complex characteristics from the images in the MRI dataset. Then, the VGG16 model is integrated with the XGBoost classifier for the classification of meningioma grades. By comparing the performance of two models, VGG16 + XGBoost model outperforms the simple CNN, with an accuracy of 98.4%, in binary classification tests. PubDate: Fri, 20 Jun 2025 22:19:29 +000
Abstract: The brain tumor grows abnormally in the human brain, which causes brain cancer. Death rates have been rising annually for the past few decades due to negligence of early treatment of brain tumors. To reduce the death rate, early identification of tumors is crucial. Early brain tumor detection may potentially lower the risk of life. Manual tumor diagnosis is complex, challenging, and time-consuming for medical professionals. Therefore, automatic detection and segmentation methods simplify the diagnosing procedure. Thus, automatic segmentation and classification methods are taken up to make the diagnosis process easy. This research proposes a two-dimensional cumulative sum average filter (2D-CSAF) for preprocessing images and an improved deviation sparse fuzzy C-means (IDSFCM) with neighbor information for segmenting brain tumors from magnetic resonance images. The novel IDSFCM segmentation increases the noise reduction capability and enhances segmentation accuracies. The hybrid modified sine cosine algorithm-crow search algorithm (MSCA-CSA)–based WELM model is proposed to classify the brain tumor. The MSCA-CSA algorithm optimizes the weights of the WELM model to increase the classification capability. The gray level co-occurrence matrix (GLCM) feature extraction technique is employed to extract the features from the segmented images, and extracted features are given as input to the MSCA-CSA-WELM model for classification. The brain tumor dataset from Harvard Medical School is considered for this research. The proposed IDSFCM segmentation achieved 99.53% segmentation accuracy. The accuracy, specificity, and sensitivity performance measures are considered for the classification. The classification performance was evaluated using accuracy, sensitivity, and specificity metrics. The proposed MSCA-CSA–based WELM model outperformed feature extraction–based classifiers, achieving 99.37% accuracy, 99.87% sensitivity, and 99.44% specificity during training. PubDate: Wed, 11 Jun 2025 01:05:09 +000
Abstract: The rise of generative artificial intelligence (AI), such as ChatGPT, enhances higher education through personalized learning, administrative automation, and increased accessibility. However, it also raises ethical concerns about data security, academic integrity, algorithmic bias, and learner autonomy. This study employs a Hybrid Thematic SWOT (HT-SWOT) analysis to examine the strengths, weaknesses, opportunities, and threats of generative AI in global higher education. Through a systematic literature review of recent studies (2020–2024), this research highlights both the benefits, such as personalized learning, accessibility, and administrative efficiency, and the risks, including digital divides, misinformation, technosolotionism, and ethical concerns. The findings emphasize the need for responsible AI policies, faculty training, and equitable implementation strategies. This study provides actionable insights for policymakers, educators, and technologists to navigate AI’s ethical integration while promoting global equity and sustainable educational practices. Addressing these challenges requires a balanced approach that safeguards academic integrity while harnessing AI’s potential to enhance education worldwide. PubDate: Sat, 24 May 2025 04:48:36 +000
Abstract: The proliferation of fake online and AI–generated news content poses a significant threat to information integrity. This work leverages advanced natural language processing, machine learning, and deep learning algorithms to effectively detect fake and AI–generated content. The utilized dataset, combined with multiple open-source datasets, comprises 43,000 real, 31,000 fake, and 80,000 AI–generated news articles and is augmented with an ensemble large language model. We combined three open-source LLMs (GPT-2, GPT-NEO, and Distil-GPT-2) into an ensemble LLM to generate new news titles, selecting the best outputs through majority voting for further dataset expansion. Preprocessing involved data cleaning, lowercasing, stop word removal, tokenization, and lemmatization. We applied six machine learning and five natural language processing models to this dataset. The two top-performing natural language–based models (RoBERTa and DeBERTa) have been combined to develop an ensemble transformer model. Among the machine learning models, random forest achieved the highest performance, with an accuracy of 92.49% and an F1 score of 92.60%. Among the natural language processing models, the ensemble transformer model attained the highest results, with 96.65% accuracy and an F1 score of 96.66%. The proposed ensemble model is optimized by applying model pruning (reducing parameters from 265M to 210M, improving training time by 25%) and dynamic quantization (reducing model size by 50%, maintaining 95.68% accuracy), enhancing scalability and efficiency while minimizing computational overhead. The DistilBERT-Student model, trained using a balanced combination of feature- and logit-based distillation from the RoBERTa-base Teacher network, achieved strong classification performance with 96.17% accuracy. Visualize-based attention maps are constructed for different news categories to enhance the interpretability of the applied transformer–based ensemble news detection models. Finally, a website was developed to enable users to identify fake, real, or AI–generated news content. The employed dataset, including AI–generated news articles and implementation scripts, can be found at the following website: https://github.com/ishraqisheree99/Combined-News-Dataset.git. PubDate: Wed, 21 May 2025 23:33:45 +000
Abstract: The focus of this paper is on the use of machine learning for the prediction of the strength outcomes of basalt fiber-reinforced concrete (BFRC), based on its mechanical properties. These target properties are compressive, flexural, and tensile strengths, estimated with knowledge of 10 variables, including cement and aggregate content, among other fiber characteristics. Models explored for regression in this paper include linear regression, K-nearest neighbors (KNN), random forest (RF), XGBoost (Extreme Gradient Boosting), support vector machine (SVM), and artificial neural networks (ANN). The highest performance among these was observed for the KNN at flexural strength with a score of 0.8737, XGBoost for compressive strength with a score of 0.8963, and RF for tensile strength with a score of 0.9420. Bayesian optimization was employed to tune hyperparameters to enhance the accuracy of the model. This study also applied Synthetic Minority Oversampling Technique (SMOTE) to generate 1000 synthetic concrete mix designs for the data to increase its diversity and allow the investigation on optimal performances regarding strength. The findings of this study contribute to advancing sustainable manufacturing practices by leveraging machine learning techniques to optimize material properties, thereby supporting the development of resilient infrastructure and enhancing industrial innovation. PubDate: Wed, 21 May 2025 05:19:15 +000
Abstract: Since the outbreak of Coronavirus Disease 2019 (COVID-19), the virus has posed a grave threat to human health. Automated segmentation of COVID-19 lung computed tomography (CT) scans is a crucial diagnostic tool that aids physicians in providing accurate and timely diagnoses, as it contains significant radiomics information. Given that the specificity for discriminating between the causes of conventional pulmonary features is lower than its sensitivity, the primary goal of this study is to develop and evaluate a CT-based radiomics model capable of distinguishing between COVID-19 and other lung diseases. To address this, we propose an efficient, modified radiomics feature processing method that integrates an optimal aerial perspective (OAP) parameter-based intensity dark channel prior (IDCP) with a 50-layer residual deep neural network (ResNet50 DNN) for autolesion segmentation (ALS-IOAP-DNN). To further enhance COVID-19 lesion estimation, novel optimization strategies, including a hybrid simulated annealing-cuckoo search (SA-CS) algorithm, are introduced alongside the original SA method. The SA-CS algorithm extends SA by preventing entrapment in local minimum and enhancing global exploration. Five benchmark functions are used to accelerate convergence and address the issue of local optima. As a result, the hybrid approach outperforms 10 recent studies on two publicly available datasets (COVID-CT-Dataset and HUST-19), achieving an average accuracy score of 100% across different epochs, along with perfect accuracy, 100% sensitivity, and 100% specificity. The proposed models significantly outperform the baseline model, with accuracy improvements of 13.6% on Data1 and 2.5% on Data2. While the baseline model achieves 88% accuracy on Data1 and 97.6% on Data2, the proposed ALS-IOAP-DNN4 model attains perfect accuracy (100%) on both datasets, demonstrating the effectiveness of ALS and advanced optimization techniques. Furthermore, the use of OAP with IDCP enhances the precision of COVID-19 lesion estimation, underscoring its significance in COVID-19 diagnosis and medical imaging management. PubDate: Mon, 19 May 2025 06:52:25 +000
Abstract: Collaborative filtering (CF)-based personalized recommendation systems are among the most widely adopted strategies for addressing information overload in modern markets. However, a significant challenge that hinders the effectiveness of these systems is data sparsity. To address this limitation, advanced methodologies such as matrix factorization and deep learning models have been developed to improve CF system performance in sparse data scenarios. Despite their advancements, most existing models rely solely on either implicit or explicit user–item interactions to generate recommendations. Traditional CF approaches often struggle with sparsity, necessitating the development of hybrid methods that integrate both explicit and implicit feedback to enhance performance. To address this gap, this study introduces DeepBlendRec, an innovative deep learning-based framework that leverages both explicit user ratings and implicit user behaviour. By utilizing this dual-input methodology, the model captures richer information regarding user–item interactions, thereby significantly improving recommendation accuracy. DeepBlendRec employs an enhanced autoencoder with a constrained decoder to process implicit ratings, thereby improving reconstruction quality and facilitating the creation of robust latent space representations. Simultaneously, explicit ratings are processed through a multilayer perceptron. The reconstructed outputs are then fused to generate a Top-N recommendation list. Experimental evaluations conducted on the MovieLens datasets demonstrate that DeepBlendRec consistently outperforms existing models across several key performance metrics, including mean-squared error (MSE), root-mean-squared error (RMSE), mean absolute error (MAE), precision, recall, and F1-score. These results highlight the potential of DeepBlendRec to advance the capabilities of recommendation systems in handling data sparsity and improving predictive accuracy. PubDate: Thu, 15 May 2025 23:05:36 +000
Abstract: Forecasting weather parameters, particularly temperature and solar radiation, plays a vital role in enhancing the efficiency of photovoltaic (PV) systems. This study introduces a cutting-edge transformer-based model specifically tailored for long-term time-series forecasting, aimed at improving the performance of PV generation systems. Leveraging the robust attention mechanisms and parallel processing capabilities inherent in encoder–decoder transformer architecture, the model effectively captures intricate relationships within weather data. A unique positional encoding layer is incorporated to bolster the model’s comprehension of the chronological sequence of data points. Furthermore, the multihead attention mechanism adeptly identifies interactions between key meteorological factors, especially temperature and radiation, which are crucial for precise PV generation predictions. Evaluation using real-world weather datasets reveals that the proposed model significantly surpasses conventional forecasting methods in mean squared error and mean absolute error metrics. This work underscores the applicability of transformer models in predicting temperature and radiation for PV generation, offering a scalable and efficient forecasting solution vital for sustainable energy management. The model is suitable for both large-scale solar installations and smaller setups, enhancing operational strategies and energy capture. Its improved accuracy in forecasting global horizontal radiation and temperature contributes to better planning and more effective energy utilization. PubDate: Fri, 09 May 2025 00:04:08 +000
Abstract: The emergence of deep learning has markedly enhanced the identification and diagnosis of ocular diseases, providing considerable benefits compared to conventional machine learning techniques. This research investigates the application of deep feature extraction for classifying eight different ocular diseases. The VGG16, a pretrained convolutional neural network (CNN) model, was employed for feature extraction, while the fine k-nearest neighbor (KNN) classifier was utilized for classification. Experimental results showed an initial classification accuracy of 89.2% using features from Gradient-weighted Class Activation Mapping (Grad-CAM) heatmaps and 83.1% using Multiscale Retinex (MSR) enhanced images. However, combining both feature sets led to an improved classification accuracy of 96.5%. Despite these promising results, several challenges remain, including the need for models that generalize across diverse patient imaging and demographics modalities, the accessibility of extensive annotated datasets, and the interpretability of models. Ethical issues and legal frameworks are also crucial for the safe and fair implementation of AI in medical services. The study suggests that future efforts in deep learning for ophthalmology should focus on creating large-scale, annotated datasets to enhance the detection of ocular diseases. PubDate: Wed, 07 May 2025 23:33:13 +000
Abstract: Retinal vessel segmentation algorithms are crucial in automated retinal disease screening systems, as accurate determination of blood vessel structures is vital for ocular disease identification and diagnosis. In this study, we propose an efficient approach combining dynamic contrast stretching (DCS) method, advanced morphological operation–based segmentation, triangle thresholding, and a final postprocessing step. Our preprocessing step enhances the contrast of retinal images acquired under various lighting conditions, enabling reliable and accurate segmentation. This enhancement is achieved using the DCS method, which is compared to two widely used contrast enhancement techniques: adaptive histogram equalization (AHE) and contrast-limited adaptive histogram equalization (CLAHE). The second step combines morphological operations and triangle thresholding to enhance vessel structures, eliminate noise, and separate blood vessels effectively. Postprocessing addresses artifacts and ambiguous areas at image boundaries. Our approach is evaluated using widely recognized reference datasets, including Digital Retinal Images for Vessel Extraction (DRIVE), Structured Analysis of the Retina (STARE), and High-Resolution Fundus (HRF). The experimental results demonstrate that the proposed method achieves superior segmentation accuracy compared to the state-of-the-art techniques. Specifically, we achieve average accuracy rates of 98.08%, 97.14%, and 98.94% for the DRIVE, STARE, and HRF datasets, respectively. In addition, our method is distinguished by exceptionally fast execution times, reaching 0.013 s for the DRIVE and STARE datasets. These results underline the importance of our time-reduced approach to improving the accuracy and efficiency of fully automated retinal disease screening systems. PubDate: Wed, 07 May 2025 23:19:23 +000
Abstract: This study introduces a novel approach to engineering design optimization through the development of an improved mountain gazelle optimizer (iMGO) that incorporates variable neighborhood search (VNS) techniques. The enhanced algorithm effectively addresses engineering optimization challenges by identifying optimal design solutions within specified constraints. In particular, iMGO significantly improves solution diversity and mitigates the risk of premature convergence to local optima, thereby overcoming the limitations of the original MGO. A comprehensive analysis was conducted using 12 functions from the CEC 2022 benchmark suite, and the algorithm was applied to five engineering problems, including the design of an I-beam, pressure vessel, three-bar truss, cantilever beam, and tension spring. Comparative results indicate that iMGO outperforms established metaheuristic techniques, such as MFO, WOA, GOA, MPA, TSO, and SCSO, as well as the original MGO. The results validate iMGO’s effectiveness in navigating the complexities of constrained engineering optimization. For instance, in practical applications, the manufacturing cost of the pressure vessel design was reduced from 6014.4537 to 5915.3358, and the weight of the tension spring was decreased from 0.0149154 to 0.0130101 relative to the original MGO. These enhancements underscore the significant potential of iMGO in real-world applications across aerospace engineering, structural design optimization, energy system planning, and other fields, thereby contributing to more efficient and sustainable engineering solutions. PubDate: Mon, 05 May 2025 22:19:10 +000
Abstract: This study proposes a multivariate time-series classification approach using deep learning to predict stocks likely to be flagged by the Market Surveillance Measure List in the Thai stock market. Formulated as a binary classification problem, the model distinguishes At-Risk and Normal stocks based on two primary datasets: End-of-Day stock prices and Market Surveillance Measure List records, incorporating trading volumes and technical indicators. To address data imbalance, concept drift, and long-term dependencies, the framework integrates feature engineering, cost-sensitive learning, and rolling window training. Experimental results show deep learning models significantly outperform traditional baseline methods in capturing financial risk patterns. The study identifies models that effectively balance predictive accuracy with computational efficiency, with performance varying based on forecasting horizons. Despite improvements from specialized techniques, the study identifies challenges in long-term financial risk prediction. These findings support market surveillance, algorithmic trading, and portfolio risk management, with future work exploring explainable AI, adaptive learning, and alternative data sources to enhance interpretability and long-term forecasting. PubDate: Fri, 02 May 2025 23:45:30 +000
Abstract: Predicting water temperature () in tropical environments is crucial for ecosystem monitoring and the sustainable management of water resources. Highly accurate and reliable forecasts are essential for the ecological management of rivers. This study evaluates the performance of machine learning-based predictive models in forecasting in the Catu River. The models were trained using climatic and hydrological data collected from 2009 to 2016 and validated with real data from 2023. The evaluated models include backpropagation neural network (BPNN), Random Forest, Bidirectional LSTM (BiLSTM), Air2Stream, and NARX, employing nine input variables such as atmospheric pressure, air temperature, and water vapor concentration. The results show that the BiLSTM model achieved the best performance, with a root mean square error (RMSE) of 0.12°C and = 0.98, followed by BPNN with an RMSE of 0.18°C and = 0.91, and the Random Forest model, which obtained an NSE of 0.95. These models demonstrated a strong ability to predict under both normal and extreme conditions, capturing the thermal dynamics of the Catu River with high precision during events involving minor thermal variations. Conversely, the NARX and Air2Stream models exhibited lower performance, proving more prone to errors under conditions of extreme variability. The findings of this study provide valuable scientific insights for river prediction and the protection of aquatic ecosystems, with practical applications in water resource management in tropical regions. PubDate: Fri, 02 May 2025 05:35:11 +000
Abstract: This study presents ensemble machine learning (ML) models for predicting residential energy consumption in South Africa. By combining the best features of individual ML models, ensemble models reduce the drawbacks of each model and improve prediction accuracy. We present four ensemble models: ensemble by averaging (EA), ensemble by stacking each estimator (ESE), ensemble by boosting (EB), and ensemble by voting estimator (EVE). These models are built on top of Random Forest (RF) and Decision Tree (DT). These base predictor models leverage historical energy consumption patterns to capture temporal intricacies, including seasonal variations and rolling averages. In addition, we employed feature engineering methodologies to further enhance their predictive abilities. The accuracy of each ensemble model was evaluated by assessing various performance indicators, including the mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination . Overall, the findings illustrate the efficiency of ensemble learning models in providing accurate predictions for residential energy consumption. This study provides valuable insights for researchers and practitioners in predicting energy consumption in residential buildings and the benefits of using ensemble learning models in the building and energy research domains. PubDate: Tue, 29 Apr 2025 20:33:42 +000
Abstract: This research focuses on developing an intelligent administrative chatbot for Roman Urdu to overcome the language barrier that hinders individuals who are not fluent in English from utilizing existing chatbot frameworks. While chatbot architectures can be rule based or artificial intelligence (AI) based, the core objective of a chatbot is to effectively address user queries across diverse domains. AI-based chatbots have demonstrated higher interactivity, scalability, adaptability to human interests, and evolving knowledge compared with rule-based architectures. However, the majority of existing chatbot frameworks is in English, which poses a challenge in regions like Pakistan where only a small percentage of the population is proficient in English. In response to this challenge, this research introduces an intelligent administrative chatbot specifically developed for Roman Urdu. Roman Urdu is the practice of writing Urdu, the native language of Pakistan, using the English alphabet. The proposed chatbot allows students to post their questions in either English or Roman Urdu, enabling them to obtain conclusive answers. The chatbot system utilizes a sentence analysis approach to generate relevant and productive responses to user queries. The core building block of the proposed chatbot is the recurrent neural networks (RNNs) sequence-to-sequence (Seq2Seq) model with long short-term memory (LSTM) units. To train the model, a dataset was meticulously collected from various administrative offices at the University of Engineering and Technology (UET), Lahore, initially in English and subsequently translated into Roman Urdu using different writing styles. The effectiveness of the proposed system was evaluated through a human judgment approach, assessing the contextual relevance and productivity of the chatbot’s responses to relevant questions. In conclusion, this research aims to bridge the language gap in chatbot frameworks, enhancing accessibility and usability. PubDate: Mon, 28 Apr 2025 22:18:26 +000
Abstract: Stroke disease has been the leading cause of death globally for the last several decades. Thus, the death rate can be decreased by early recognition of disease and ongoing surveillance. However, the largest obstacle to perform advanced analytics using the conventional approach is the growth of massive amount of data from various sources, including patient histories, wearable sensor devices, and medical data. The current technology that could have a large impact on the healthcare sector is the integration of machine learning with big data analytics (scalable machine learning), particularly in the early diagnosis of this disease. To address this issue, a scalable stroke disease prediction model for a multinode distributed environment, which was developed by combining big data analytics concepts with machine learning to handle extensive healthcare datasets, an aspect not seen in the prior literature on stroke disease detection, is presented in this work. We have implemented four scalable algorithms: logistic regression, random forest, gradient-boosting tree, and decision tree, using a dataset that was collected from a Medical Quality Improvement Consortium database. As a result, two worker nodes and one master node were used to analyze the dataset. The model’s performance was assessed using performance metrics including the area under the curve (AUC) and confusion matrix. With an accuracy of 94.3% and an AUC score of 99%, the random forest was determined to be better based on the experimental results. It was also shown that the main risk factor for stroke disease is diabetes, which is followed by hypertension. This study demonstrated the effectiveness of using Spark’s scalable machine learning techniques to forecast stroke disease and identify risk factors earlier. The findings of this study can be utilized by physicians as clinical decision aids to aid in the more accurate identification of stroke disease. PubDate: Mon, 21 Apr 2025 23:35:12 +000
Abstract: In this paper, a cooperative hunting optimal control problem is studied for multimarine surface vehicle (MSV) systems based on reinforcement learning (RL). First, in order to enhance the efficiency of cooperative hunting, a novel task allocation method is proposed based on a leader–follower structure, where follower and leader MSVs perform the target encircling task and the target tracking task, respectively. Second, on the basis of task allocation, adaptive control design is employed for the follower MSVs to enhance the control performance of target encircling; optimal feedback control design combined with adaptive feedforward control design is considered using the RL algorithm for the leader MSVs to ensure the optimality of target tracking. Finally, the stability of the multi-MSV hunting control system is guaranteed and all signals are uniformly ultimately bounded based on the Lyapunov theory in the closed-loop system. The effectiveness of the proposed scheme is demonstrated through simulation results. PubDate: Mon, 21 Apr 2025 20:49:44 +000
Abstract: Computer vision plays a crucial role in current day tasks due to its wide range of merits and applications in many fields like disease diagnosis, medical decisions, military, security, scientific applications, and business. Identifying plant species upon their leaf images is a very challenging and complex task. In this study, we introduced a model for classifying a variety set of plant leaves using different techniques such as factorization machine (FM), dimensionality reduction (DR), and ensemble learning (EL). In our study, FM is used as a classification algorithm, whereas DR is included in two stages, namely, feature extraction (FE) and feature selection (FS). FE has been used to extract the features from images in tabular form; on the other hand, FS is used to reduce the size of the data by getting rid of noisy (redundant and irrelevant) features. In addition, we utilized two methods for FS, filter-based and wrapper-based approaches. The used filter was the minimum redundancy maximum relevance, and the used wrapper was the improved binary cuckoo search algorithm that is based on Lévy flight and abandon nest functions. Regarding EL, it was used to declare different versions of FM by passing different subsets of features that are deduced from the original ones to improve its performance. We used the Swedish and Flavia leaf datasets for training and testing phases. The proposed model achieved a high performance in terms of accuracy that produced 95.67% on Swedish dataset and 99.6% on Flavia dataset. According to the results and in comparison to other methods, we proved that our proposed model ensures a favorable plant leaf classification approach. Finally, the proposed wrapper FS approach showed very good results without setting up the number of features to be selected by the user. The entire implementation of this work can be found at https://github.com/Mohammed-Ryiad-Eiadeh/Binary-Cuckoo-Search-For-Plant-Leaf-Prediction. PubDate: Tue, 15 Apr 2025 00:46:39 +000
Abstract: Social media sites facilitate users’ discussions, expression of opinions, sharing of information and news, and promotion of ideas and products, thus rapidly increasing the volume of hate speech and offensive content on online platforms. Consequently, hate speech and offensive content have turned out to be a widespread issue that negatively affects both individuals and society and needs to be controlled through detection and removal. This paper aims to provide a further understanding of the meaning of hate speech and offensive language and to provide a comprehensive discussion of the various techniques to detect hate speech and offensive language. A systematic literature review (SLR) of 90 research papers published between 2018 and 2024 was conducted to discover gaps within the literature. This review revealed challenges and possibilities for further development and improvement of previous findings. The results show that most of these works classified hate speech and offensive language using techniques from machine learning (ML) and deep learning (DL), and the most common performance metrics were Accuracy, Precision, Recall, and F-measure. The benchmark datasets are also described. Twitter was the most commonly utilized social network for obtaining datasets, while Facebook is sometimes used. Moreover, the findings of this review offer insight into research trends in Arabic hate speech and offensive language, as well as new research directions. The most interesting finding is that until now most Social Media Network Developers have not included autodetection of hate speech or offensive language plugins. Finally, this study presents a guideline for choosing the best strategies and techniques to detect and predict Arabic offensive language and hate speech. PubDate: Sun, 13 Apr 2025 20:36:15 +000
Abstract: Hospital readmissions impose a significant financial strain on healthcare systems and can adversely affect patients. Unfortunately, traditional approaches to predicting readmissions frequently lack accuracy. This presents a critical challenge, as identifying patients at high risk for readmission is essential for implementing preventive measures. The study introduces a novel method that employs machine learning to automatically extract features from patient data, eliminating labor-intensive manual feature engineering. The primary goal is to develop predictive models for unplanned readmissions for UTI patients at Jordan University Hospital within 3 months postdischarge. This is executed through a retrospective analysis of electronic health records from January 2020 to June 2023. By leveraging machine learning techniques, the study identifies high-risk patients by evaluating demographic, clinical, and outcome characteristics, ensuring model reliability through thorough optimization, validation, and performance assessment. Three predictive models were developed as follows: a gradient-boosting classifier (GBC), logistic regression (LR), and stochastic gradient descent (SGD). The GBC, SGD, and LR achieved impressive accuracy rates of 99%, 95%, and 89%, providing strong confidence in the methodology. The study’s findings reveal key risk factors associated with readmissions, enhancing our understanding of this process and offering a valuable framework for improving patient care, optimizing resource allocation, and supporting evidence-based decision-making in healthcare management. PubDate: Thu, 10 Apr 2025 05:31:56 +000
Abstract: In the rapidly evolving landscape of decision-making, particularly amid the complexities of managing extensive data, it is crucial to develop methodologies that facilitate informed choices. This paper presents a novel technique designed to address multicriteria decision-making (MCDM) challenges through the opportunity losses-based polar coordinate distance (OPLO-POCOD) method. By integrating the principle of opportunity losses—a fundamental concept in economics and management—into the evaluation process, this approach enhances the decision-making framework, providing a clearer understanding of the trade-offs involved in selecting alternatives. The innovative methodology is applied to assess the sustainability impact of Industry 4.0 technologies within Iran’s automotive industry, a sector facing unique challenges such as the need for technological adaptation. Utilizing the fuzzy OPLO-POCOD method, the study analyzes various sustainable performance indicators, ultimately identifying simulation, blockchain, and radio frequency identification as critical enablers of sustainable development. These technologies exhibit remarkably low opportunity loss scores of 0.0212, 0.0232, and 0.0270, respectively, underscoring their pivotal roles in enhancing energy and resource efficiency in production activities. This research provides significant insights for industry stakeholders and policymakers, emphasizing the importance of integrating advanced technological solutions into sustainable development strategies. By aligning technological advancements with sustainability objectives, this study not only illuminates the transformative potential of Industry 4.0 technologies but also paves the way for future inquiries in this essential field, offering a strategic roadmap for promoting sustainable practices in the automotive sector and beyond. PubDate: Mon, 07 Apr 2025 20:55:16 +000
Abstract: A vast variety of neural network (NN)–based controllers use indirect adaptive control structures for their implementation, which primarily aims at estimating the nonlinear dynamics of the system and thereby generating a suitable control action. However, the most commonly used gradient descent weight update rule of the NN-based indirect adaptive controllers often fails to identify the system dynamics appropriately, and the control action becomes ineffective on the system under control. Further, a significant number of existing works in this domain employ randomized or suboptimal initial NN weights, which can potentially hamper the transient performance of the control loop. To address these issues, this paper proposes an innovative control scheme that utilizes the strength of indirect adaptive control of NN, along with the robustness of the PID controller. Since the PID control structure is often independent of plant dynamics and generates a control action to mitigate any error between the reference and the current plant output, it can be easily augmented with the control action generated by inverse neural network (INN) to mitigate any effects of unlearnt dynamics by INN. Further, we have used a bio-inspired optimization algorithm, that is, particle swarm optimization (PSO), to optimize the initial weights of the INN along with the PID controller’s parameters to get an optimal control performance. The proposed INN + PID controller scheme has been tested on a cart-mounted inverted pendulum system due to its challenging control requirement owing to its intricate nonlinear dynamics. Detailed simulation studies for the proposed INN + PID and PID controllers have been carried out for various control requirements, viz. set point tracking, disturbance rejection, and robustness testing. Further, an extensive comparative study has been devised based on the integral of absolute error (IAE) to test the efficacy of the proposed INN + PID controller against the conventional PID controller. Through extensive comparative studies, it was deduced that the proposed INN + PID controller is capable of handling the intricate nonlinear dynamics of the cart-mounted inverted pendulum system and provides a sturdy stabilization of the angular position of the pendulum with respect to the desired trajectory and superior transient control in comparison with the conventional PID controller. In terms of quantitative comparison, the improvement in IAE achieved by the proposed INN + PID controller was found to be 94.84%, 94.62%, and 69.86% better in comparison to the conventional PID controller for set point tracking, disturbance rejection for introduced impulsive force, and time-varying force variation, respectively. PubDate: Fri, 04 Apr 2025 03:59:08 +000
Abstract: Skin cancer spreads quickly as the skin is the most vulnerable organ, and melanoma (MEL) is a fatal type of skin cancer. Detecting MEL in the early stage can hugely increase the chance of a cure. There are several methods based on machine learning to detect MEL from dermoscopic images. However, increasing the accuracy of detection is still challenging. This paper presents a new method for MEL detection that considers the combination of deep and handcrafted time–frequency local features. After short preprocessing, the convolutional neural networks (CNNs) extract the deep features. To this end, feature maps at the output of the flatten layer are considered as deep features. The scale-invariant feature transform (SIFT) descriptors are handcrafted local features computed from the four subbands of one-level two-dimensional discrete wavelet transform (2D DWT). After the fusion of the mentioned features, semisupervised discriminant analysis (SDA) reduces the highly correlated and redundant features. The Bayesian optimizer finds the optimum parameters of the SDA and Gaussian kernel of the support vector machine (SVM) classifier to maximize the classification accuracy. The HAM10000 dataset with data augmentation is considered to assess the performance of the proposed method. Simulation results show that the proposed method reaches the accuracy and sensitivity of 94.19% and 96.22%, respectively. The most challenging parts of the proposed method are extraction of deep features and tuning the parameters of SDA and Gaussian-SVM. PubDate: Mon, 31 Mar 2025 06:49:14 +000
Abstract: A neurodevelopmental illness called autism spectrum disorder (ASD) is frequently associated with sensory problems including an excessive or insufficient sensitivity to noises, scents, or touch. Children with autism typically do not talk much and keep to themselves, but they can imitate certain actions from cartoons and movies. They may exhibit unsafe or unexpected behavior as a result. Early detection and therapy can assist in improving the diseases. In this study, we suggested a data-driven machine learning (ML) model for examining the autism dataset of diverse age groups (toddlers, children, and adults) to classify autism in the initial stage. The proposed ML model can efficiently analyze autism patients’ datasets and correctly classify and detect ASD features. We utilized a data preprocessing technique followed by feature selection methods using information gain and Pearson correlation. Then, we employed five different ML classifiers (KNN, RF, SVM, NB, and MLP) together with a hyperparameter optimization strategy. We assess their work using performance metrics such as prediction accuracy and the F1-measure. After comparing the accuracy between different classifiers, SVM produced the highest accuracies of 98%, 99%, and 100% for the toddler, child, and adult datasets while MLP produced an accuracy of 0.94 for the Pakistani child dataset, respectively. These thorough experimental assessments suggest that correct fine-tuning of the ML techniques can be crucial in the classification of autism in individuals of various ages. We believe that the thorough feature significance analysis presented in this study can guide medical professionals’ judgment when screening ASD individuals. In comparison with other methods currently used for the timely identification of ASD, the suggested framework has shown encouraging results. PubDate: Tue, 25 Mar 2025 04:22:38 +000
Abstract: Handling missing values presents a critical challenge in thyroid disease prediction, significantly impacting diagnostic accuracy. This study evaluates the effectiveness of cold-deck, mean, and K-nearest neighbor (KNN) imputation techniques for predicting thyroid disease using a dataset of 9172 observations with 31 clinical features (5.2% missing values). Feature importance analysis identified thyroid-stimulating hormone (TSH), thyroxine (TT4), and free thyroxine index (FTI) as consistently significant biomarkers across all imputation methods. Five classifiers—Naïve Bayes, linear regression, support vector machines (SVM), LightGBM, and recurrent neural networks (RNN)—were assessed on imputed datasets, with performance evaluated through accuracy, F1 score, and recall. The KNN imputation method enhanced LightGBM’s accuracy by 0.47% over mean imputation (99.06% vs. 98.99%) and by 1.47% over cold deck (99.06% vs. 98.59%), demonstrating its superiority in preserving feature relationships and enhancing predictive power. LightGBM achieved the highest performance with KNN imputation (accuracy: 99.06%, F1: 97.57%, and recall: 97.83%), outperforming other classifiers by 2.5%–4.0% in accuracy. These results underscore the necessity of robust imputation techniques for reliable thyroid disease prediction. The study provides a reproducible framework for managing missing data in healthcare analytics, emphasizing the interplay between imputation, feature importance, and classifier selection to optimize diagnostic accuracy. PubDate: Mon, 24 Mar 2025 05:34:22 +000